morningman opened a new pull request, #67789:
URL: https://github.com/apache/doris/pull/67789
### What problem does this PR solve?
Problem Summary:
Nothing in the tree asserts the bytes a MySQL client actually receives, or
the Arrow batches an
Arrow Flight SQL client actually receives. Every test around the frontend's
result path stops at the
`QueryState` or at a `ShowResultSet`; the packet framing below that -- how
many packets, in what
order, with which sequence ids, terminated by an EOF or an OK, flushed where
-- is only exercised by
real clients in the regression suites, and only for the handful of
statements those suites run.
That gap is about to matter. The session and result layers are being
reorganized so that MySQL and
Arrow Flight SQL become two front ends over one session (the follow-up PRs
extract a protocol
adapter and a result sender out of `ConnectContext`, `ConnectProcessor` and
`StmtExecutor`). Those
are pure refactors, and a pure refactor that drops an EOF packet, reorders a
column definition, or
shifts a sequence id passes every existing unit test and shows up only as a
client that hangs.
This PR takes the baseline first, before anything moves. It adds no
production code.
### What is changed?
Four test classes under
`fe/fe-core/src/test/java/org/apache/doris/qe/protocol/` and two golden
files under `fe/fe-core/src/test/resources/protocol-golden/`:
- `RecordingMysqlChannel` -- a `DummyMysqlChannel` that plays canned request
packets and records
every response packet: its payload, the sequence id it was framed with,
and whether the response
was flushed at it. The sequence id is advanced the way the real channel
advances it (once per
packet read, once per packet written), so what is recorded is what would
reach the wire.
- `ProtocolGolden` -- renders the recorded traffic as an annotated hexdump
and compares it with the
checked-in file. On a mismatch it writes the current traffic to
`target/protocol-golden/<name>`
and names it in the failure message, the same convention
`AccessControlBehaviorBaselineTest` uses.
- `MysqlPacketGoldenTest` -- 27 cases: constant selects, a NULL literal, an
empty result set,
`SHOW VARIABLES`, `DESC` (the one case whose row stream is longer than a
single packet), `SET`,
`USE`, `EXPLAIN`, a syntax error, an unknown table, an error followed by a
healthy statement on the
same connection, multi-statement requests with and without
`CLIENT_MULTI_STATEMENTS`, and the
connection commands (`COM_FIELD_LIST`, `COM_STMT_PREPARE`,
`COM_STMT_CLOSE`, `COM_SET_OPTION`,
`COM_RESET_CONNECTION`, `COM_PING`, `COM_INIT_DB`, `COM_STATISTICS`, an
unknown command,
`COM_QUIT`). Two of them run with `CLIENT_DEPRECATE_EOF` negotiated off,
the capability that
decides whether a result set ends in an EOF or an OK and whether the
column definitions get their
own terminator.
- `FlightResultGoldenTest` -- the same statements an Arrow Flight SQL
session can be answered
frontend-side, recorded as the schema and rows of the `VectorSchemaRoot`
the session caches, or as
the error it failed with.
Capability negotiation is replayed rather than mocked: the test derives the
effective capability the
way `MysqlProto.negotiate()` does (`server & client`) and sets
`clientDeprecatedEOF`,
`clientMultiStatements` and the serializer capability from it, so the
recorded packets match what a
real connection with those flags would produce.
Two responses are recorded as shape rather than bytes, because their payload
legitimately moves with
changes that have nothing to do with the protocol: a parser error carries
the entire keyword list of
the grammar (4881 bytes today), and an `EXPLAIN` carries the current plan
text. For those the golden
keeps the packet kinds, plus the error code, SQL state and the start of the
message.
The statement set is limited to what a frontend answers on its own.
`COM_STMT_EXECUTE` is
deliberately absent: `ConnectContext.supportHandleByFe()` is false for it,
so its result always comes
from a backend and belongs in `prepared_stmt_p0`, not here. No case in
either test reaches a backend
(verified: the mocked backend logs no `exec_plan_fragment` request during
the run).
The two golden files are added to `.licenserc.yaml`'s `paths-ignore`: both
tests regenerate the file
in full and compare it byte for byte, so a license header would be read back
as unexpected content.
### Release note
None.
### Check List (For Author)
- Test
- [x] Unit Test
- [ ] Regression test
- [ ] Manual test (add detailed scripts or steps below)
- [ ] No need to test or manual test. Explain why:
- Behavior changed:
- [x] No.
- [ ] Yes.
- Does this need documentation?
- [x] No.
- [ ] Yes.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
